README.md

Lesson 83: Synthetic Panoptic Segmentation (Toy, CPU-friendly)

This lesson adds a minimal panoptic segmentation training loop on synthetic scenes.

• synthetic data with stuff background + thing rectangles
• semantic labels plus instance masks/classes per sample
• model wired to the local dlhub.vision.panoptic_segmentation family via build_panoptic_fpn_panoptic_segmenter
• compact loss: semantic CE + instance class CE + mask BCE

Run:

python -m tracks.vision.lesson_83_synthetic_panoptic_segmentation.train \
  --epochs 1 \
  --max-train-batches 2 \
  --max-eval-batches 1 \
  --device cpu